Sam Stephens

Socialist, music/art/culture, intellectual property/creative commons/etc, software development.

#norobots

Location: Patea, New Zealand
Pronouns: https://pronoun.is/he/him
Music: https://postmoderncore.com

Last Notes

2026-08-20 01:13:42 UTC

- reply

I won't touch them for now, because I think there's profound ethical issues with most of the big LLMs, and because I want a view over a much longer time period of how LLMs work in software development before I change an approach I've spent more than 25 years building, that works very well for me.
In particular, how do things shake out after the bubble bursts, once you're paying the real costs of the LLM you're using?

2026-08-20 00:56:52 UTC

- reply

I agree, but lots of people think they are because they can get good results right now, and do not understand how that doesn't mean this approach will continue to work over longer time frames.

2026-08-20 00:44:45 UTC

- reply

no code runs in my coffee maker, I use a plunger/french press, but I hear what you're saying.
I guess I bang on about this because a lot of the people who are really excited about agentic code generation are contributing to code bases that will never be finished, and I think a lot of people in industry do not understand that there's a big gap between writing software that works now, and software that can continue to be developed effectively throughout it's lifetime.

2026-08-14 00:56:08 UTC

- reply

>That argument presupposes that software can't be simply finished. That a codebase is ever-evolving, never done.
Fair, but I suspect not many codebases are ever finished; only code built for certain purposes can be finished. I've been in business software for most of my life (digressing through AWS), and no codebase I ever worked with in business finished, because businesses are always evolving, and requirements are always changing.
I think my thoughts apply to most codebases.

2026-08-14 00:50:09 UTC

- reply

Also, unless LLMs are carefully managed, they appear to introduce bloat and redundancies into codebases over time.
So my conviction is that between the human side and the code side, that long term ownership of a codebase is basically impossible, and that if the bubble collapsing doesn't destroy agentic coding, that we'll start to see codebases end up at the point where an LLM cannot successfully make changes to it, and a human cannot understand it to solve that problem.

2026-08-14 00:49:16 UTC

- reply

I've read plenty of anecdotal evidence of people finding their diligence declining over extended periods of agentic coding.
Then from the other side, just because an agentic system can be used to create a functioning system, I'm not convinced they can own it for any length of time. The problem as I see it is thinking of the LLMs interactions with a codebase in information theoretical terms; each interaction an LLM has with a codebase will introduce noise to the signal.

2026-08-14 00:48:35 UTC

- reply

Firstly from the human diligence side, I'm convinced it's not actually possible to maintain agentic coding over long periods of time and keep up high standards, due diligence, and ensuring understanding of your code. As an agentic coder, you're what Cory Doctorow calls a reverse centaur, a machine assisted by a human being, who has the impossible task of monitoring every line the agent produces when faults are rare.

2026-08-14 00:48:15 UTC

- reply

when you say "It demonstrably can work", for me, I think it depends on your definition of work. For me, building a piece of software is the easy bit; the hard bit is owning it, fixing it, and evolving it over a time scale of years. I'm convinced that no-one is going to manage to do this with agentic coding.

2026-05-23 22:44:13 UTC

- reply

I think what we're going to see is that LLM codebases will end up at a point where the LLM can no longer successfully modify them, and are complex and convoluted enough to defy human understanding (with the exception of people who are being careful about the code structure, but you cannot do that, and simultaneously generate too much code to review). A similar danger to that of humans writing code without architectural vision, but amplified by the volume of code an LLM generates.

2026-05-23 22:42:18 UTC

- reply

"They can’t even do human code reviews anymore as there’s too much code being generated" this is the ticking time bomb. I've been saying for a long time that writing code is easy; it's owning code that's hard. Designing an application to change and grow as future requirements arrive takes deliberate design thought. LLMs are not capable of that.

2024-12-23 08:32:38 UTC

- reply

The Governess for sure. Anne Hegerty is a legend, so funny and so smart. And autistic too. And the best chaser according to some random website.